Artificial Neural Network For Recognition Of Handwritten Devanagari Character

IOSR Journal of Computer Engineering (IOSR-JCE)
e-ISSN: 2278-0661,p-ISSN: 2278-8727, Volume 17, Issue 1, Ver. I (Jan – Feb. 2015), PP 60-64
www.iosrjournals.org
DOI: 10.9790/0661-17116064 www.iosrjournals.org 60 | Page
Artificial Neural Network For Recognition Of Handwritten
Devanagari Character
Ms. Seema A. Dongare1
, Prof. Dhananjay B. Kshirsagar2
, Mr. Nitin J.Khapale3
,
Ms. Amruta D. Pawar4
12
(Computer Department, SRESCOE Kopargaon, India)
3
(E&TC Department, KBP Polytechnic Kopargaon, India)
4
(IT Department, KBP Polytechnic Kopargaon, India)
Abstract: Information gathering is one of the most important thing in the digital world, many organizations
still use handwritten documents for storing their information. So it is important task to convert this information
in to digital world. Handwritten character recognition helps user to convert their handwritten documents in to
digital information. Purpose of this dissertation work is to implement zone based feature extraction method
which is combination of image centroid zone and zone centroid zone of individual character image. In feature
extraction using zone based approach, individual character image is divided into n equal size zones then
average distance of all pixels with respect to image centroid or zone centroid is computed. In combination of
image centroid and zone centroid approach it computes average distance of all pixels present in each zone with
respect to image centroid as well as zone centroid which gives feature vector of size 2Xn features. This feature
vector is presented to feed forward neural network for recognition.
Keywords: Feed forward neural network, handwritten character recognition, image centroid zone, zone
centroid zone.
I. Introduction
Optical character recognition converts scanned images of printed or handwritten text into digital text.
Handwritten character recognition is a field of OCR. Basically there are two classes of HCR as off line character
recognition and on line character recognition. In off line character recognition, writing is captured optically by
scanner while in on line character recognition coordinates of successive points are represented as function of
time as well strokes made by user are also considered. Handwritten character recognition is an intelligent optical
character recognition capable of handling the complexity of writing, writing environment, materials, etc.
Handwritten Devanagari characters are quite complex for recognition due to presence of header line,
conjunct characters and similarity in shapes of multiple characters. The main purpose of this paper is to
introduce a method for recognition of handwritten Devanagari characters using segmentation and neural
networks. The whole process of recognition works in stages as preprocessing on document image, segmentation
of document into lines, line into words and word into characters, finally recognition using feed forward
neural network. Important steps in any HCR are preprocessing, segmentation, feature extraction and recognition
using neural network [2].
II. Related Work
K. Y. Rajput and S. Mishra have proposed a system for recognizing handwritten Indian Devanagari
script. In feature extraction character matrix as an array of black and white pixels of size 30X30 is prepared.
Afterwards, the Feed Forward neural network with back propagation is used in learning and recognition process.
[3].
S. Arora, D. Bhattacharjee proposed two stage classification approaches for handwritten Devanagari
characters. The first stage is using structural properties like detection of shirorekha, spine in character and
second stage exploits some intersection features of characters which are presented to a feed forward neural
network. Each handwritten character can be adequately represented within 16 segments (each of size 25 X 25
pixels) and hence 32 features for each character can be used as input to neural network [4].
V. Agnihotri proposed Handwritten Devanagari script recognition using neural network. Diagonal
based feature extraction is used for extracting features of the handwritten Devanagari script. These feature set is
converted into chromosome bit string of length 378.Individual character image of size 90x60 pixels is divided
into 54 equal sized zones. Each zone has 19 diagonal lines and the foreground pixels present along each
diagonal line is summed to get a single sub feature, thus 19 sub features are obtained from each zone.
These 19 sub features values are averaged to form a single feature value and placed in the corresponding
zone. Finally, 54 features are extracted for each character [5].

Artificial Neural Network for Recognition of Handwritten Devanagari Character
D. Singh, S. Singh and Dr. M. Dutta proposed twelve directional feature inputs depending upon the
gradients. This technique can recognize all types of handwritten characters even special characters in any
language [6].N. Sharma, U. Pal, F. Kimura, and S. Pal have proposed a quadratic classifier based scheme for the
recognition of offline Devanagari handwritten characters. Features used in the classifier are obtained from the
directional chain code information of the contour points of the characters. This technique has achieved 98.86%
and 80.36% recognition accuracy on Devanagari numerals and characters, respectively [7].
III. Devanagari Script
3.1 Properties of Devanagari Script
Devanagari script has features different from other languages. Devanagari character set has 13 vowels,
36 consonants and 10 numerals with optional modifier symbols. Characters are organized into three zones as
upper, middle and lower zone. Core characters are positioned in middle zone, while optional modifiers in upper
and lower zones. Two characters may be connected to each other. In Devanagari script, the concept of
uppercase and lowercase characters is absent. Fig. 2 represents Devanagari character set. It represents
Devanagari character modifier set. Modifiers are optional symbols arranged in upper and lower zones.
3.2 Issues Regarding Recognition of Devanagari Script
Some reasons that cause recognition of Devanagari characters difficult are as:
1. In Devanagari Script individual characters are connected by header line (Shirorekha) which makes
segmentation of individual character is quite difficult.
2. Characters may be connected to form conjuncts for which separation is complex.
3. Presence of modifiers makes segmentation difficult.
4. Some Devanagari characters are similar in shape.
Fig. 1 Devanagari Character Set
IV. Proposed System
HCR works in stages as preprocessing, segmentation, feature extraction and recognition using neural
network. Preprocessing includes series of operations to be carried out on document image to make it ready for
segmentation. During segmentation the document image is segmented into individual character or numeric
image then feature extraction technique is applied on character image. Finally feature vector is presented to the
neural network for recognition.
Fig. 2 Block Diagram of system

4.1 Preprocessing
The preprocessing consists of series of operations as grayscale conversion, noise removal, and
binarization. After selecting Devanagari document image, color image is converted into gray scale. Unwanted
contents are removed from image. Then binarization is applied on gray scale image.
4.2 Segmentation
Once Image preprocessing is done it is necessary to segment document into lines, line into words and
word into characters. When individual character has been separated from document we can extract features
from it for recognition.
4.3 Feature Extraction
For feature extraction we have implemented grid based approach which is the combination of
image centroid zone and grid centroid zone of individual character or numerical image. In this
technique individual character image is resized to a size of 256X256 and then divided into 16 equal
sized grids each of size 64X64. Average distance of all pixels with respect to image centroid and zone
centroid is computed. It gives feature vector of size 2X16 features.
Algorithm: Image Centroid Zone (ICZ) and zone centroid zone (ZCZ) feature extraction.
Input: Preprocessed individual character image.
Output: Extracted features for classification and recognition.
Algorithm:
Begins
Step 1: Divide an input image in to n equal sized grids.
Step 2: Compute distance between the image centroid and each pixel present in the grid.
Step 3: Compute the distance between the grid centroid and each pixel present in the grid.
Step 4: Repeat step 2 and 3 for the entire pixels present in the grid.
Step 5: Computation of average distance between these points with respect to image centroid.
Step 6: Computation of average distance between these points with respect to grid centroid.
Step 7: Repeat this procedure for all grids.
Step 8: Obtaining 2Xn such features for classification and recognition.
Ends
Fig. 3 Feature Extraction from Devanagari Numeral Image “A”
Fig. 3 shows an illustration of procedure to extract features from Devanagari character "A". Let
d1,d2,d3,...,dn are distances from image centroid similarly D1,D2,D3,...,Dn are distances from the zone
centroid, then compute average distances between these points separately. This gives 2 feature values for each
zone. Same procedure is repeated sequentially for each of zone. With combination of ICZ and ZCZ we will have
two feature values per zone which gives 32 feature values provided no of zones are as 16 [1].

4.3.4 Recognition of Characters using Neural Network
The back end used for performing recognition is neural network. In the off-line recognition system, the
neural network is fast and reliable tool in order to achieve high recognition accuracy. This module will
implement Artificial Neural Network using error back propagation (EBP) algorithm.
Fig. 4 shows structure of feed forward neural network with one input, hidden and one output layer. We have
implemented feed forward neural network with following specification.
 Input nodes: 32 input neurons (2 X 16 for combination of ICZ and ZCZ)
 Output nodes : 41 output neurons.
 No of Hidden layer : 1 with 4 nodes.
 Training algorithm : Error Back Propagation.
 Performance function : Mean Square Error.
Fig. 4 Neural Network Structure
V. Result And Discussion
We have studied various research papers which reveal that the difficulty of handwritten character
recognition has two aspects. The first is attributed to the writer variations in style, size, shape, ink color, ink
flow and thickness, digitization imperfections etc. The second is the deficiencies of the particular method used
for feature extraction. With consideration of these two aspects neural network training and testing has been
performed and observations are recorded.
We have created synthetic dataset consisting of total 2460 patterns out of which we have used 1640
patterns for training and 820 patterns for test. We have divided this dataset into two datasets as Dataset1 and
Dataset2 which contains 820 and 1640 patterns of each alphabet respectively. Testing has been performed on
separate dataset containing 20 patterns of each alphabet total of 820 patterns. During testing we have recorded
accuracy in terms of number of characters recognized accurately with percentage accuracy for each character.
From Table1 we can analyze that with increase in size of dataset used to train neural network, accuracy
of recognition increases. Accuracy of the system depends on writing order and style, font size and color,
thickness.
Table1. Combined Result Set for Dataset1 and Dataset2
Sr. No. Character % Accuracy
Training: Dataset1
% Accuracy
Training: Dataset2
1 k 70 80
2 ba 55 80
3 [ 55 80
4 na 55 80
5 t 55 75
From these two we can observe that when size of training dataset increases recognition accuracy
increases. Neural network accuracy depends on size of dataset used to train it. Larger the dataset increases
accuracy but decreases speed of recognition. For few characters accuracy remains same or slight increase has
been observed, reasons behind this can be due to similarity in shapes of multiple Devanagari characters, or may
be simply due to variations in writing styles. From Fig. 5 we can conclude with increase in size of dataset used
to train neural network, accuracy of recognition increases.

Fig. 5 Comparative Accuracy Analysis for Dataset1 and Dataset2
The dataset consists 2460 patterns of Devanagari alphabets. We have used zone based feature
extraction techniques on these samples in which we have extracted 32 feature values from each character image.
We have obtained different recognition accuracy on different characters. First we have tested system on 20 test
patterns of each alphabets, the highest accuracy obtained for character is 90. Then we have tested it on test
document with alphabets here maximum accuracy we have achieved is 77. And finally testing has been done on
different documents where we get average accuracy as 75%.
VI. Conclusion
Development of handwritten Devanagari OCR is a challenging task in Pattern recognition area. Here,
we proposed a method which does the segmentation of handwritten document into lines, words and characters.
And further recognition process will be done with the help of neural network. This system needs to be tested on
a wider variety of images containing characters in diverse fonts and sizes. It has been found that recognition of
handwritten Devanagari characters is quite difficult task due to presence of header line, conjunct characters and
similarity in shapes for multiple characters.
Acknowledgement
I am very much thankful to my respected project guide and Head of Dept. Prof. D. B. Kshirsagar,
for his ideas and help proved to be valuable and helpful during the creation of this dissertation work. I am
also thankful to our P.G. Coordinator Prof. P. N. Kalavadekar, for helping m e while selecting and
preparing dissertation work. I would like to thank all the faculties who have helped me during my
dissertation work. Lastly, I am thankful to my friends who shared their knowledge in this field with
me.
References
[1]. G. Sinha, Mrs. R. Rani, Prof. R. Dhir ,Recognition Using Zonal Based Feature Extraction Method and SVM Classifier ,
International Journal of Advanced Research in Computer Science and Software Engineering , ISSN: 2277 128X, Volume 2,
Issue 6, June 2012.
[2]. M. Patil, V. Narawade, Recognition of Handwritten Devanagari Characters through Segmentation and Artificial neural networks ,
International Journal of Engineering Research and Technology (IJERT) , ISSN:2278-0181,Vol. 1 Issue 6, August 2012.
[3]. K. Y. Rajput, S. Mishra, Recognition and Editing of Devanagari Handwriting Using Neural Network, Proceedings of SPIT-IEEE
Colloquium and International Conference, Vol. 1.
[4]. S. Arora, D. Bhattacharjee, M. Nasipuri, L. Malik and B. Portier, A Two Stage Classification Approach for Handwritten
Devanagari Characters.
[5]. V. Agnihotri, Offline Handwritten Devanagari Script Recognition, IJITCS, 2012, 8, 37-42.
[6]. D. Singh, S. Singh and Dr. M. Dutta, Handwritten Character Recognition Using Twelve Directional Feature Input and Neural
Network, International Journal of Computer Applications, 0975-8887, Vol. 1.
[7]. N. Sharma, U. Pal, F. Kimura and S. Pal, Recognition of Off-Line Handwritten Devanagari Characters Using Quadratic Classifier.
[8]. V. Bansal, R. Sinha, Segmentation of touching and fused Devanagari characters , Pattern Recognition 35,875-893, 2002.

Artificial Neural Network For Recognition Of Handwritten Devanagari Character

More Related Content

What's hot (19)

Similar to Artificial Neural Network For Recognition Of Handwritten Devanagari Character (20)

More from IOSR Journals (20)

Recently uploaded (20)

Artificial Neural Network For Recognition Of Handwritten Devanagari Character